Female-RHINO: A Real-Time Scanner-Integrated Framework for Automated Quantitative Uterine MRI Analysis and Structured Reporting

Technical Summary: Female-RHINO Framework

Problem Statement
Standardized assessment of uterine MRI remains difficult due to significant anatomical variability, observer dependence, and the absence of workflow-integrated automated analysis tools. Current clinical practice relies on manual, post-examination quantitative assessment, which is time-consuming, subject to inter-observer variability, and often documented in narrative reports that hinder standardization and longitudinal comparability. Furthermore, conventional MRI workflows lack interactive quantitative feedback during the scan, limiting the ability to adapt imaging strategies in real-time. While deep learning methods have shown promise for segmentation and landmark detection, most existing approaches are offline, task-specific, and not integrated into the clinical acquisition workflow.

Methodology
The authors present Female-RHINO ((R)eproductive (H)ealth (I)maging A(N)alysis T(O)ol), an end-to-end, real-time AI framework integrated directly with the MRI scanner. The system operates on sagittal T2-weighted pelvic MRI volumes and consists of the following components:

• Data and Training: The system was trained and evaluated on over 500 multi-center datasets spanning diverse protocols, vendors (Philips, Siemens), field strengths (0.55T to 3.0T), and patient populations. The datasets included segmentation labels for the uterine wall, uterine cavity, uterine myomas, and Nabothian cysts, as well as anatomical landmarks.
• Automated Segmentation and Volumetry: A 3D full-resolution Residual Encoder UNet (ResEnc-UNet-M) implemented via the nnUNet-v2 framework performs automated segmentation. The model uses z-score normalization and patch-based training with combined Dice and cross-entropy loss. Post-processing involves 3D connected component analysis to identify and enumerate distinct lesions (myomas and Nabothian cysts) for volumetric calculation.
• Automated Landmark Detection and Biometry: A Swin-UNETR network detects six anatomical landmarks (fundus, endometrial cavity apex, internal os, external os, and two anteroposterior diameter points) within a cropped region of interest (ROI). Landmarks are localized by regressing 3D Gaussian heatmaps. Derived biometric measurements include uterine body length, cervical length, anteroposterior diameter, fundal thickness, utero-cervical angle, and uterine position (anteverted/retroverted).
• Real-Time Integration and Reporting: The framework connects to the MRI scanner via the Fire interface. During acquisition, DICOM slices are streamed in real-time, converted to NIfTI format, and processed by the GPU server. The system generates a structured, clinician-oriented HTML report containing segmentation overlays, biometric measurements, volumetry, pathology counts, and visualizations within approximately 70 seconds of scan completion.

Key Contributions

1. Scanner-Integrated Real-Time Analysis: Unlike offline tools, Female-RHINO operates inline with the MRI acquisition, enabling immediate availability of quantitative results during the ongoing scan.
2. End-to-End Automated Workflow: The system unifies segmentation, volumetry, lesion characterization, and landmark-based biometry into a single pipeline without manual interaction.
3. Structured Reporting: It replaces narrative reporting with standardized, quantitative reports that include visual verification overlays and reference ranges for contextual comparison.
4. Robustness Across Heterogeneity: The framework was validated across three distinct protocols with varying field strengths, vendors, and acquisition parameters, demonstrating generalizability.

Results
Evaluation on independent retrospective and prospective cohorts yielded the following performance metrics:

• Segmentation: Mean Dice Similarity Coefficients (DSC) were 0.82 for the uterus and 0.80 for fibroids. Nabothian cyst segmentation showed lower agreement (DSC ~0.30–0.61 depending on the protocol) due to their small size and variability.
• Landmark Detection: The system achieved a mean radial error (MRE) of 3.7 mm across all landmarks. The failure rate (defined as error > 10 mm) was low, with 6 failures out of 270 total landmarks evaluated.
• Prospective Validation: In a prospective deployment (n=5), automated volumetry showed strong agreement with manual measurements (DSC 0.85 vs. inter-observer 0.83). Biometric measurements showed mean absolute differences of 2.33 mm and 2.20 mm compared to manual observers, comparable to inter-observer variability (2.01 mm).
• Efficiency: End-to-end processing, from image acquisition to report generation, was completed in under 70 seconds.

Significance and Claims
The paper claims that Female-RHINO enables real-time, scanner-integrated AI for automated uterine MRI analysis. The authors assert that this system has the potential to improve standardization, efficiency, and clinical workflow in pelvic imaging by providing immediate, reproducible, and structured quantitative data. By integrating analysis directly into the acquisition workflow, the framework supports adaptive imaging strategies and reduces the reliance on time-consuming manual post-processing. The authors position this as a step toward "workflow-aware and adaptive pelvic MRI," where quantitative AI analysis can directly support image acquisition planning and clinical interpretation, particularly for evaluating uterine anomalies and reproductive health. The study acknowledges limitations, including the lack of layer-specific segmentation (e.g., junctional zone) and challenges with landmark localization in cases of severe anatomical distortion, but maintains that the framework offers a robust foundation for automated gynecological imaging.